AISTATS Artifact Evaluation

Use this for evidence packaging around AISTATS. The venue centers on artificial intelligence, statistics, and machine learning, so artifacts should make statistical and computational claims inspectable.

Artifact plan

• Decide what evidence reviewers need: proof details, derivations, simulation scripts,

benchmark code, datasets, preprocessing, hyperparameter sweeps, random seeds, logs, or qualitative examples.

• Keep decision-critical evidence in the main paper or appendix; optional run files can live

in supplementary material.

• Anonymize repository history, paths, notebook metadata, license headers, organization

names, cluster paths, grants, and commit authors.

• Include a minimal reproduction map: environment, dependencies, hardware, commands, expected

outputs, runtime, seeds, and known nondeterminism.

• For restricted data, give enough provenance and processing detail for credible

reproduction without violating data-use terms.

• After acceptance, replace anonymous archives with public, licensed, citable artifacts when

feasible.

What AISTATS evidence reviewers open first

| Claim type | First artifact inspected | Common failure caught | |---|---|---| | Convergence rate or regret bound | Proof appendix and constants | Condition used in the proof but missing from the theorem statement | | Monte Carlo simulation | Seeded simulation script | Plots cannot be regenerated because seeds and replication counts are absent | | Benchmark comparison | Training and evaluation configs | Baseline tuning budget undocumented | | Bayesian or MCMC method | Sampler diagnostics and chain logs | No convergence statistics or trace evidence anywhere |

Because AISTATS reviewers are often statisticians, they will rerun a small simulation far more readily than they will retrain a deep model, so make synthetic studies turnkey before polishing anything else.

Worked vignette: packaging a Monte Carlo study

A hypothetical submission proposes a doubly robust treatment-effect estimator with a root-n normality guarantee, validated on synthetic causal data plus two real benchmarks.

• Ship the data-generating process as one parameterized script rather than constants buried

in notebooks, so reviewers can vary n, dimension, and confounding strength.

• Record the replication count and the exact seed sequence used for every coverage and bias

table; AISTATS-style claims about interval coverage are meaningless without them.

• Emit tables directly from logged results so the PDF numbers and artifact numbers cannot

drift apart.

• State explicitly where the simulated regime satisfies the theorem assumptions and where it

deliberately violates them, since that mapping is what statistical reviewers grade.

Calibration anchors

• Supplementary inspection at AISTATS is at reviewer discretion; assume only the README and

one entry script get opened, and design accordingly.

• Upload size limits and accepted formats vary by cycle; verify against the current

OpenReview submission form rather than past years.

Output format

text
[Artifact role] anonymous supplement / camera-ready release / public archive
[Contents] <code/data/proofs/logs/notebooks>
[Anonymity risks] <paths/metadata/licenses/URLs>
[Reproduction level] turnkey / scripted / descriptive / weak
[Fixes before upload] <ordered list>